Appendix C — Explaining This to Someone Who Is Not Technical

This is not a summary of the book. It is a translation layer for one situation: you are in a room with a senior officer who did not come up through technology, who has limited time, and who has already been pitched AI by at least three vendors.

Read it in the car.

The four sentences

Everything in this book compresses to four sentences. If you get through only these, you have done the job.

1. The cases are already detected. They are not connected. About a thousand new online fraud cases a day are reported in Thailand. Every one is already known. Nobody is failing to detect crime. What is missing is the link between one victim's report and the other three hundred reports belonging to the same operation.

2. The link is in the words, and the words are not in any database. The account numbers are in fields, and the banks already share those through the Central Fraud Registry. What sits in no field is the victim's own account of what happened — what they were told, who contacted them, what the script was. That is the record of the human operation, and it is sitting in prose that nothing can search.

3. Reading that prose at scale is the one thing a machine is genuinely good at. Not judging. Not deciding. Reading a hundred thousand narratives and reporting which ones resemble each other, with the file numbers attached so a person can check.

4. It runs on your own machine, and the data does not leave. No cloud. No foreign vendor holding case material. That is a constraint, and it is also the reason this can start next week instead of after a procurement.

Stop there. Four sentences is enough for a first meeting and every extra one costs you.

The two questions you will be asked

"We already have that."

You probably will be told this, and there is a good chance something adjacent is true — the Central Fraud Registry is real and operating, and a THB 200 million AI system for AOC was reported as targeting a Q3 2026 launch. Appendix B has the details.

The correct response is not to argue. It is one question:

"Does it read the free-text account of what the victim was told, or does it work on the account numbers?"

If the answer is account numbers, you have found your gap and you did not have to claim anything. If the answer is that it reads the narrative too, then the work is already being done and you should say so, thank them, and ask what would still help. An honest exit from a room is worth more than a manufactured disagreement, and people remember which one you did.

"How is this different from what a big vendor sells?"

Three real differences, in the order they will land:

Language. Established link-analysis tools connect structured fields — account numbers, phone numbers. Reading a Thai-language narrative account and extracting what happened from it is a different task, and it is not what those tools were built for.

Time and cost. A procurement runs in years. A demonstration on synthetic data runs in a week.

Where the data sits. On the unit's own hardware, with nothing leaving the building.

And the sentence that should follow immediately, because it is true and because it is what a sceptical officer is waiting to hear: this is a layer on top of what you already have, not a replacement for it. Nobody in that room wants to hear that the system they signed off on is wrong.

What to say about the risk, before you are asked

Raise this yourself. It is counterintuitive and it works, because everyone in the room has read about AI going wrong and is waiting to find out whether you know.

"There are documented cases where this technology has caused serious harm. In the United States, at least four people were wrongly arrested on facial recognition matches. Every one of those cases has the same cause: the computer's answer was treated as evidence instead of as a lead. So the rule I would build in is that the system may only produce pointers — every output is a list of case numbers a human opens and checks. It never asserts anything."

That paragraph does more for your credibility than any capability claim, for a specific reason: it demonstrates that you have thought about how the thing fails, which is what a person who is accountable for outcomes actually needs to know about you.

What to say about Singapore, honestly

Singapore's Anti-Scam Centre is the right comparator for this region and it will resonate. Use it, with the qualifier attached:

"Singapore built exactly this and reports recovering more than seven hundred million dollars since 2019. I should say that those are the police force's own figures and nobody outside has audited them — but it is a working model with a public track record, and Thailand's problem is larger than theirs."

The qualifier is not weakness. The officer across the table already assumes every number he is shown has been polished. Being the person who says so first makes everything else you say more credible.

What not to say

Do not quote a single national loss figure as settled. The sources genuinely disagree — one implies about 62 million baht a day, another says 70; full-year 2025 appears as both THB 115.3bn and about 89bn. Quote a case count with its window instead, or say that the figures differ and name both. Someone in that room may know the discrepancy.

Do not say the police cannot solve these cases. You do not know their clearance rate. Nobody does — it is not published. Asserting a failure you cannot measure, to people who live inside the work, is the fastest way to lose a room.

Do not promise a timeline for anything beyond a demonstration.

Do not offer to look at real case data in that meeting, and do not accept it if offered informally. Every reason is in Chapters 8 and 9 — lawful basis, chain of custody, the PDPA question nobody has answered for individuals. A synthetic-data demonstration is not a limitation you are apologising for. It is evidence that you understand the rules.

The question to ask before you leave

One question, and it is for you rather than for them:

"If this worked, which budget would it sit in?"

An answer means there is a path. No answer means it was a conversation. Both are useful to know, and you can only find out by asking while you are still in the room.

The one-page leave-behind

If you leave anything, leave this. Not a deck.

The problem. ~1,000 online fraud cases reported daily in Thailand; 121,921 cases and THB 7.48bn in losses in the first 120 days of 2026 (ACSC, 11 May 2026). Cases are detected. Networks are not assembled.

The gap. Bank account data is already linked nationally through the Central Fraud Registry. Victim narrative — what the caller said, the script, the story — is unstructured and unsearched.

The proposal. A local, offline system that reads Thai-language complaint narratives, extracts the entities inside them, and groups reports that belong to the same operation. Output is a list of case numbers with the matching text quoted, for an investigator to open.

The safeguards. Runs on unit hardware; no data leaves. Produces pointers, never assertions. Every output resolves to a source document. Demonstration on synthetic data only.

The precedent. Singapore's Anti-Scam Centre, operating since June 2019 (police self-reported figures). Crime linkage is an established discipline in UK and US policing.

The ask. One week, synthetic data, no access to any real case file, no cost.